Hindawi Complexity Volume 2020, Article ID 8877161, 15 pages https://doi.org/10.1155/2020/8877161

Research Article Analysis Technology of Sports Match Based on Data Mining and Image Feature Retrieval

Hong Huang and Risheng Deng

School of Sports Science, Lingnan Normal University, Zhanjiang, Guangdong 524048, China

Correspondence should be addressed to Risheng Deng; [email protected]

Received 3 September 2020; Revised 25 September 2020; Accepted 30 September 2020; Published 14 October 2020

Academic Editor: Chuan Lin

Copyright © 2020 Hong Huang and Risheng Deng. )is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Tennis game technical analysis is affected by factors such as complex background and on-site noise, which will lead to certain deviations in the results, and it is difficult to obtain scientific and effective tennis technical training strategies through a few game videos. In order to improve the performance of tennis game technical analysis, based on machine learning algorithms, this paper combines image analysis to identify athletes’ movement characteristics and image feature recognition processing with image recognition technology, realizes real-time tracking of athletes’ dynamic characteristics, and records technical characteristics. Moreover, this paper combines data mining technology to obtain effective data from massive video and image data, uses mathematical statistics and data mining technology for data processing, and scientifically analyzes tennis game technology with the support of ergonomics. In addition, this paper designs a controlled experiment to verify the technical analysis effect of the tennis match and the performance of the model itself. )e research results show that the model constructed in this paper has certain practical effects and can be applied to actual competitions.

1. Introduction the game can obtain and find more valuable information and laws in a timely and accurate manner to make up for the Compared with traditional, technical, and tactical analysis deficiencies of traditional statistical methods. )erefore, it is methods, data mining technology can more clearly describe necessary for tennis players to use data mining technology to and analyze the process and sequence of each hitting assist them in making technical and tactical decisions [2]. technique, such as the position and route of the shot, and the Compared with the popularity of established sports such process and sequence of winning and losing points. as football and basketball, tennis has relatively few followers. Moreover, it can help coaches to better grasp the com- However, with the development of mass sports and the monalities and characteristics of players’ skills and tactics, influence of international tennis, tennis has developed and provide reference and basis for the use of skills and rapidly in my country in recent years. Moreover, more and tactics in future training and games [1]. In recent years, with more people have begun to pay attention to large-scale the increase of opportunities for domestic players to par- tennis matches, and the annual four matches ticipate in the competition, the increasingly abundant game and various masters have attracted the attention of many videos have made the accumulation of technical and tactical fans [3]. )e Tennis Open has introduced “Eagle Eye” data in the game continue to increase. How to explore the technology, which can accurately determine whether the ball technical and tactical characteristics of excellent players is out of bounds. Players have two applications for the right from the rich data and provide a basis for the scientific to use Hawkeye in each set. After the referee accepts the decision-making of coaches and athletes has become one of application, they will broadcast the “Instant Replay” on the the key and urgent problems in tennis theoretical research. big screen to present the results of Hawkeye’s calculation. In addition, the use of data mining technology to analyze the )e system takes no more than 10 seconds from data col- technical and tactical characteristics and laws of athletes in lection to result presentation. However, the system is 2 Complexity composed of 8 or 10 high-speed cameras, four computers develop the best type should be players with comprehensive and a large screen, which makes the cost very high, and the skills. For this reason, coaches and athletes should ensure amount of calculations when processing the collected data is that training sessions and personal training include five very large [4]. Based on this, this article uses big data actual combat training items: serving, receiving, bottom technology to analyze tennis sports matches, combined with counter attack, cross-ball, and online . )e literature machine vision to identify sports scenes, and improve the [12] believed that most players play more defensively than analysis effect of tennis sports matches. offensively during the game. Even so, the result is that the mistakes in the game are greater than the active scoring, and 2. Related Work in the game, the risk of offense is greater than that of defense. )e literature [13] studied the differences in the tactics used )e literature [5] pointed out that when analyzing players by all-round players in the face of different play types. In the and correcting their movements, a correct understanding of face of serving and Internet players, all-round players often biomechanics will help coaches avoid focusing on peculiar, have to improve their first success rate; that is, it does seemingly uncomfortable shots and help them focus on the not hurt to spin more and slow down when serving. timing of shots. )e literature [6] proposed that the sequence Moreover, by changing the serve , all-round players can of tennis coordination chain body segments is from the foot, serve the opponent’s body more often, thus avoiding the hip, upper body, arm or shoulder, and elbow to wrist, and opponent’s direct attack when receiving the serve. )e lit- biomechanics is from knee flexion, hip rotation, body ro- erature [14] mentioned that the two-handed tation, arm around shoulder, elbow extension, and forearm stroke technique is the backhand stroke technique used by internal rotation to wrist rotation. )e literature [7] analyzed most athletes. Some athletes who are born with height and the throw technique of tennis players in the Atlanta Olympic strength conditions are not particularly dominant, but have Games and showed that the throwing position is in front of good flexibility and fast speed, and can use the two-handed the front foot or left front. High-speed camera technology is backhand technique to better develop their tennis potential. used for player analysis. )e results show that many ex- At the same time, it recommends that players strengthen cellent players start to hit the ball after the ball has started to their physical fitness while learning to use the two-handed fall (25–20 cm). backhand technique, and continuously improve their tennis )e literature [8] mentioned technology alone cannot IQ, so that they can master the technique more quickly and win the game; even if the technology is improved in daily promote the popularization of tennis. )e literature [15] training, if it cannot be played in the actual game, it is divided the technique of receiving and serving into two meaningless. At the same time, it is also mentioned that sections in the research of receiving and serving tennis. tennis is essentially control and anticontrol, and if the athlete Among them, one is the front section of the shot, and the controls the opponent throughout the game, he will defi- other is the shot section. )e front part of the shot is nitely be able to win the game. Moreover, it is not only composed of three parts: , position, and body posture. superlevel performance that can succeed. If the opponent )e position is generally near the bottom line and close to can barely deal with it, the game will inevitably develop the sideline of singles, but the position should be selected towards the side that is beneficial to oneself, and then victory according to the serving power and spin of the server and can be expected. )e literature [9] mentioned that in today’s one serve and two serve. )e literature [16] used the women’s games, offense after serving is a very common knowledge of system theory to conduct in-depth research on tactic. )e sooner the athletes play, the better. Moreover, the tennis serve when exploring the tennis serve system, and best measure is to start within the bottom line, which can put believed that tennis serve is not just a technical action, but a more pressure on the opponent. After serving, the backhand complete system. )e literature [17] pointed out that the attacking athletes use the backhand to grab high points from speed of serving is an effective way to affect the opponent’s the baseline, which sometimes poses a greater threat to the offensiveness and direct scoring. Under the condition of a opponent than hitting hard. )at is, it is commonly referred certain speed, the player’s choice of the ball’s landing point to as turning back. )e literature [10] analyzed the im- can avoid the opponent’s special skills and attack the op- pression of female tennis players from physical, technical, ponent’s weakness, which can more effectively restrain the and psychological factors, and elaborated on the charac- opponent’s performance and win the victory of the serve. teristics and trend distribution of the playing styles in )e literature [18] used the American Peak three-dimen- modern female tennis. Moreover, it proposed that today’s sional high-speed video system and the video analysis system female tennis players have great offensive skills, and there of the Aijie Human Information Institute to measure and will be more female players adopting more active offensive analyze the kinematics data of the strong serve technique of styles. For this reason, coaches should focus on developing domestic outstanding female tennis players. In the end, it is young tennis players in the most ideal way and be fully concluded that the vigorous serve of tennis is a whipping prepared for future offensive play. )e literature [11] believes action. Moreover, the basic principles of the technique of that in tennis, players tend to adopt a certain style of play serving are proposed and demonstrated. In addition, it is based on their own unique technical, physical, tactical, and proposes that the hitting action is not a natural transition psychological characteristics, and imitating a style of play at and extension of the swing action, but a relatively inde- a young age also has an impact on the type of play used by an pendent action technique with its own technical charac- athlete. Moreover, it believes that players who want to teristics. )e literature [19] took high-quality tennis matches Complexity 3

→ as the research object. )e success rate, scoring rate, is an indicator function; when o ( x ) � c, its value is l, →k straightness rate, and the outcome of a high-quality tennis otherwise it is 0. An input vector x is given, and the output match are related as follows. )ose with a high success rate prediction classification mark y of the random forest → H often cannot win the game, while those with a high scoring model H( x ) is [22] rate have a very high probability of winning. )e literature → [20] concluded that the first serve success rate and the first yH � arg max p(y � c | x ). ( ) c∈D 4 serve win rate in a tennis match are not positively correlated, Y and the statistical value of the first serve success rate has no )e random forest model modeling algorithm generally substantial significance, and its level has little effect on the uses the following steps: for k � 1, ... , k, outcome of the game. However, the first serve win rate and n ( n N) the second serve win rate have a greater impact on the (1) k 0 < k < data are sampled from the data set outcome of the game, and they are the two key indicators with replacement. that determine the outcome of the game. From the above (2) According to the nk data, a decision tree without analysis, it can be seen that the relevant technology for tennis pruning is trained. When building a decision tree match analysis is currently relatively vacant. )erefore, this model, the attribute set used to create a branch node article combines the actual needs of the tennis match to is to randomly select a part of all attributes instead of construct a corresponding visual analysis model, and per- using the entire attribute set to create a branch node. forms action feature recognition and strategy improvement )e output of most random forest models adopts a with the support of big data technology. special form of formula (1) that is, we set wk � 1/k. )erefore, formula (1) is simplified to [23]

3. Machine Learning Model k )is paper proposes three methods of action knowledge → 1 → H( x ) � � o ( x ). (5) extraction based on the random forest model, so the first step k k k�1 is to establish a random forest model. In the process of establishing the random forest model, a data set (X, Y) is given. Among them, X � �x1, ... , xN � is the attribute vector Definition 1. Attribute segmentation: a random forest al- 1 N → set and Y � �y , ... , y � is the classification label set. Each gorithm model H( x ) will be given, and each attribute xi, i � →i i l attribute vector x � �x1, ... , xM � has M attribute values; 1, ... ,M will be divided into a certain number of intervals among them, each attribute value xj ∈ Dj has a finite value [24]: range Dj. xj can be categorical or numerical. )e classifi- (1) If xi is categorized, and it contains n categories, then cation label Y has a finite value range DY. If there is no →i i i xi is logically divided into n intervals. ambiguity, x � �x1, ... , xM � will also be represented by → (2) If x is numerical, if it is assumed that the branch x � �x1, ... , xM � below. In order to simplify the descrip- i tion, this paper first only considers the case where the node in each decision tree in the random forest prediction classification is binary; that is, y ∈ {}0, 1 . Only algorithm model is xi > b, then b is a dividing point of small changes are needed to extend the binary classification attribute xl. If there are n division points for attribute situation to a multiclass model [21]. xl in all decision trees, then attribute xl is divided A random forest model generally consists of K decision into n + 1 intervals. → trees. Each decision tree k has an output function ok( x ), x )e number of segmentation intervals of attribute →i is which means that a classification mark y ∈ Dy is output expressed by ni, so naturally, the segmentation vector p � when the input attribute vector is given. )erefore, the (p , ... , pM) corresponds to the attribute vector output of the entire random forest is → 1 → x � (x1, ... , xM). Among them, the value of p is the index value of the segmentation interval where the attribute x is k → l → → located. Because the attribute values of the p vector and the H( x ) � � w o ( x ). (1) → k k x vector correspond to each other, it is convenient for the k�1 → x following description. In the absence of ambiguity,→ will Among them, w ∈ R is the weight of each decision tree. still be used to represent the segmentation vector p below → k In addition, H( x ) also has another probability represen- [25]. tation. A classification mark c ∈ DY is given: Definition 2. Attribute change: a random forest algorithm k → → → � w I o ( x ) � c � model H( x ) is given, and τ represents attribute change, p(y � c | x ) � k�1 k k . (2) k which is defined as a triplet τ � (xl, p, q). Among them, �k�1wd xi, i � 1, ... ,M, is a certain attribute, and the values of p and Among them, q represent two of xl attributes. If and only if there is an attribute value xi of the i-th attribute in the vector in the → interval P, the attribute change τ is executable in the given I ok( x ) � c �, (3) → vector x . )e function of the attribute change τ is to 4 Complexity

complete the transfer of the attribute value xi of the vector node set, which is a set of all possible states. A represents the from the interval p to the interval q. set (a1, a2, ... , an) of actions related to solving the problem, which can be used to transform one state into another state. Definition 3. Action: a change set of attributes is defined as G represents the set of all possible goal states related to action a, a � �τ1, ... , τ|a| �, and there is τ|a| ≥ 1, which means solving the problem. that at least one attribute change is in one action. )e )e model of Markov decision process is shown in → condition that an action executable on the x vector needs to Figure 1, which is generally represented by the five-tuple meet is that all attribute changes in the action can be exe- 〈S, A, T, R, c〉, where S represents a finite set of all possible → cuted on the vector x . )e execution of action a will states of an agent or environment. Each state in the set generate a cost value; that is, π(a) > 0. represents a possible state, and each state must contain all relevant signs for making correct decisions; A represents a → For the given attribute vector x and the action a that can limited set of all possible actions of an agent; T represents → be performed on each attribute of x , the final result vector of the state transition function T(s, a, s′), which is used to → executing action a is x ⊗ a. express the probability of performing the action a in the → In the absence of ambiguity, p � (y � c | x ) is simply state S to reach the state s′. )e Markov property is reflected → described as p( x ) [26]. in the state transition function. )e future state of the decision is only related to the current state of the decision Definition 4. Action knowledge extraction problem: the and the currently selected action. )e above content can be AKE (actionable knowledge extraction) problem can be formally expressed as →I represented by a four-tuple � � (H, x , c, A), and the H in the four-tuple represents the random forest algorithm P s � s′ | s , a , s , a , ... , s , a � � P s � s′ | s , a �. → t+1 t t t−1 t−1 0 0 t+1 t t model. At the same time, x represents the input vector of (11) the problem, c ∈ DY represents the final desired classifica- tion label, and A is the action set. Finding an action sequence R represents the reward function, and R(s, a) represents As is the ultimate goal of the AKE problem; among them, the immediate reward value that the agent can obtain by there is As � �a1, ... , an �, al ∈ A. )e constraints to be performing action a in state S. )e reward function can be satisfied by the action sequence As are related to the subsequent state, that is, R(s, a, s′). Under this definition, all s′ of R(s, a, s′) can be expected to be converted min FAs � � � πai �. A ∈A (6) into R(s, a). c is the discount factor, and its value range can s a A l∈ s be expressed as c ∈ [0, 1]. )e reason for setting the discount )erefore, we obtain factor is to make the long-term return of the infinite number →g of steps converge, which makes the Markov decision process p� x � ≥ z. (7) model have strong practical significance. When human beings make decisions, most of them pay more attention to Among them, z represents a constant, 0 < z ≤ 1, and the →g the value of returns in a short period of time, thus down- target vector x is playing the final long-term returns. When the c value is close →g →I x � x ⊗ a ⊗ ... ⊗ a . (8) to 1, it means that the agent values long-term returns. When 1 n c is close to 0, it means that the agent pays more attention to )at is, the action sequence A � �a , ... , a � is se- the immediate report. )e Markov decision process model s →I1 n quentially applied to the input vector x , and the result can control the impact of current decisions on the future by vector is finally obtained. setting the size of c. State is an ordered set of a minimum of variables )e classic planning problem can be described as a triple form D � < I, A, G > . Among them, I represents the initial q0, q1, ... , qn introduced to describe the difference between a certain type of different things, and its vector form is as state set, A represents the action set, and G represents the follows: target state set. In this type of problem, the initial state, T characteristic action, and target state are all required to be Q ��q0, q1, ... , qn� . (9) clearly stated. )e action that can change the current state of the system is an action, and the action sequence represents q � (i � , , ... , n) Each element l 0 1 in the formula is a the process evolution of the system state, as shown in component of the set, which is called a state variable. Given a Figure 2. set of values for each component, a specific state is obtained: Q ��q , q , ... , q �T. (10) k 0k 1k nk 4. Markov Decision Process to Solve the )e means of changing a problem from one state to Optimization Problem of Action another is called an action, which is generally represented by a. Knowledge Extraction )e state space of a problem (state space) is a graph that represents all possible states and their relationships to solve Reinforcement learning learns decision-making strategies the problem. )e state space is generally a directed graph, through autonomous interaction with the environment, so denoted by G � (V, A, G). Among them, V represents a that the long-term cumulative reward value received by the Complexity 5

action set, and the action is represented by a; Action T: S × A × S ⟶ [0, 1], as the state transition probability matrix, represents the probability distribution of tran- Surroundings sitioning to another state after performing an action in a certain state; R represents the reward function: a Rs � ER� t+1 | St � s, At � a�. (12) Current Reward )e above formula represents the immediate reward state value made by the environment when the state transition occurs; c represents the discount factor, and c ∈ [0, 1]. )e size of c indicates the degree of inclination to the present and future rewards. )e introduction of c is the uncertainty of the Agent future, which can avoid the infinite return of the Markov process. When c is close to 0, it is more inclined to the immediate reward value of the current time step. When c is Figure 1: Markov decision process. close to 1, it is more inclined to the long-term reward value in the future. )e reward function R is the reward signal obtained during the interaction with the environment. )e reward function reflects the nature of the current task and also Initial state serves as the basis for strategy modification. )e reward function R is to evaluate the quality of the generated action:

Rs � ER�t+1 | St � s �Gt. (13)

We assume that Gt represents the return value of all Action 1 Action 3 discount rewards from time step t: ∞ k Gt � Rt+1 + cRt+2 + ... , � � c Rt+k+1. (14) k�0 Here, the discount factor is used to control the return Action 1 k value Gt to be limited, and c R is the reward value R ob- tained through the influence of the k + 1 time step discount. For each state S in the state set S, the strategy π needs to complete an action a in the action set A; that is, it specifies the set of actions to be taken in each possible state. Strategy π : S A Goal state is the mapping from state to action, that is, π ⟶ . In the Markov decision process, our goal is to find an optimal strategy π∗ with the largest cumulative reward R : Figure π 2: Classical planning problem description. ∗ π � arg max Rπ, π ∞ (15) strategy is the largest. )e difference between reinforcement R � E ⎡⎣� ctr ⎤⎦. learning and other machine learning methods is that it needs π π t t�0 to learn strategies through interaction with the environment. ∗ Secondly, after an action is performed in the environment, if Among them, π is the optimal strategy, rt is the im- there is no mark on the quality of the action, and only after a mediate reward for actions performed at time step t, and Rπ period of interaction, the accumulated reward can be known, is the cumulative reward for actions performed at time step t so as to infer the quality of the previous action. )is is a under strategy π. At the same time, ct is the discount factor forward-looking process. )e idea of Markov decision- c ∈ [0, 1] to the power of t, and Eπ[·] is the expectation making process is used, and the expected target state can be under strategy π. obtained by determining the change of state attributes For any strategy, the expected return of long-term ac- through the reward function. Finally, the action knowledge cumulated rewards obtained by executing the strategy is is extracted, and the action knowledge can be transformed used to evaluate the pros and cons of the strategy. )e re- into intuitive recommendations for users. ward function R is an instant evaluation of a certain state (or )e most basic model of Markov decision process is a action), but the value function considers the quality of a five-tuple �M DP � �S, A, T, R, c �; among them, S represents certain state (or state-action pair) from a long-term per- a finite set of states, that is, the state space, and the state is spective, so the value function is usually called evaluation represented by s; A represents the action space of a limited function. 6 Complexity

)e expectation of accumulative rewards obtained by According to Bellman’s optimal equation, there are executing action at and subsequent strategy π in state st is ∗ ∗ V (s) � max E� rt+1 + cV st+1 � | st � s, at � a�, denoted as V(st); that is, a ∗ (20) ∞ � max � T s, a, s′ ��R(s, a) + cV s′ ��. ⎛⎝ t ⎞⎠ a V st � � E � c rt+1 . (16) s′ t�0 Among them, V∗(s′) is the optimal state value function Among them, rt is the immediate reward of the action in the next state s′. performed at time t, and the discount factor c(c ∈ [0, 1]) Algorithm performs strategy iteration: the algorithm balances the current immediate reward value with the future first randomly initializes a strategy πt and calculates the state reward value and ensures the convergence of V(s). value function vt under this strategy. According to these For any strategy π, the value function is defined to ac- state value functions, a new strategy πt+1 is obtained, and the cumulate the expected value of the discount reward value; value function vt+1 of each state under the new strategy is that is, calculated until convergence. Calculating the value of each ∞ state under a strategy is called strategy evaluation. A new π ⎡⎣ t ⎤⎦ V (s) � Eπ � c rt | s0 � s . (17) strategy based on the state value is called strategy t�0 improvement. Algorithm performs strategy evaluation: according to Among them, s represents the state of time step t, and t Bellman’s equation, the value function of a state is related to t � 0 represents that s corresponds to the initial state. 0 the value function of its subsequent states. )erefore, we use )e purpose of the reinforcement learning algorithm is the subsequent state value function v(s′) to update the to find a strategy π, which is the ultimate goal of rein- current state value function v(s). Strategy evaluation tra- forcement learning, so that the value Vπ(s) of each state S verses all states and updates its state value function reaches the maximum at the same time, namely, according to the following formula: Vπ(s) � E� r + cr + ... + ci− 1r + ... | s � s�, 1 2 i 0 Vπt (s) � � π(s, a)(R(s, a)) + c � T Vπt+1 s′ �. (18) ()s,a,s′ V∗(s) � max Vπ(s). a∈A s′ π (21) Among them, rt is the immediate reward at time step t, and c(c ∈ [0, 1]) is the discount factor. V∗(s) is called the Algorithm performs strategy improvement: the new optimal value function, and the corresponding optimal strategy is obtained according to the state value function, strategy is and the new strategy is better than the old strategy. For a state S, the strategy chooses an action a that maximizes the ∗ π π π � arg max V (s). ( ) current state value function R(s, a) + c � T(s,a,s′)V (s′). π 19 )at is, s′

⎧⎪ ⎪ ⎛⎝ πt+1 ⎞⎠ ⎪ 1 a � arg max R(s,a) + c � T V s′ � , ⎪ a ()s,a,s′ ⎨ s′ π � (22) t+1 ⎪ ⎪ ⎪ ⎛⎝ πt+1 ⎞⎠ ⎪ 0 a ≠ arg max R(s,a) + c � T V s′ � . ⎩ a ()s,a,s′ s′

)e termination conditions of the strategy iteration are

, S , ⎨⎧⎪ 0 If state is the target state πt V (s) � πt+1 (23) ⎩⎪ R(s,a) + c � T()s,a,s′ V s′ �, otherwise. s′

→ Input: �AKE � �H, x , c, A � Repeat Output: a t In the current state st, action at is selected according to Strategy π(s, a), state st, action at, reward rt im- the greedy strategy. mediately, and transition probability T(s,a,s′) is Action at is executed, and the next round of state st+1 and initialized. immediate reward rt are obtained. Complexity 7

)e algorithm first performs a strategy evaluation: )en, the algorithm performs strategy selection:

∗ ⎛⎝ πt+1 ⎞⎠ V (s) � � π(s, a) R(s,a) + c � T()s,a,s′ V s′ � . a∈A s′ (24)

⎧⎪ ⎪ ⎛⎝ πt+1 ⎞⎠ ⎪ 1 a � arg max R(s,a) + c � T V s′ � , ⎪ a ()s,a,s′ ⎨ s′ π � (25) t+1 ⎪ ⎪ ⎪ ⎛⎝ πt+1 ⎞⎠ ⎪ 0 a ≠ arg max R(s,a) + c � T V s′ � . ⎩ a ()s,a,s′ s′

Until the termination condition is met, the algorithm exits:

, S , ⎨⎧⎪ 0 If state is the target state πt V (s) � πt+1 (26) ⎩⎪ R(s,a) + c � T()s,a,s′ V s′ �, otherwise. s′

5. Tennis Video Image Processing Based on obvious parts of the image according to certain pixel Image Processing values and image grayscale and other reference factors, and at the same time removing blurry im- )e research objects and data analysis in this article are all ages. )e threshold method is usually used to seg- from professional tennis players’ matches and training, ment the image, and the specific operation steps are which are stored in the form of video images. Due to the divided into the following three steps. )e first step is complicated background of tennis competition and training, to traverse the pixel values in the image and find the it is difficult to identify the characteristics of tennis. )is point M0(x0, y0) with the largest pixel value. )e article combines image processing technology to design an second step is to combine the actual quality of the automatic processing system for tennis video images. )e image, artificially set the size of the image with 4 ∗ 3 core of the system is the TS201 processor. )is processor has pixels, and compare the maximum pixel point good image signal processing capabilities and is a widely M0(x0, y0) up, down, left, and right. Pixel compares used image processor. )e composition of the whole system the size of the found pixel value with the current includes image processor TS201, detector, storage device, maximum pixel value and finds the point M1(x1, y1) RS422 serial port device, and so on. Figure 3 shows the with the smallest pixel value. architecture of the image automatic processing system. )e third step is to obtain the difference between the )e workflow of the tennis video image automatic maximum pixel value and the minimum pixel value processing system is as follows: the high-resolution camera in the image. )e formula is as follows: collects the image signal; after the image is filtered, it is sent to the image processor TS201 through the RS422 serial port i � M0x0, y0 � − M1x1, y1 �. (27) and the signal transmission link; the feature recognition of the signal is completed in the image processor, the feature (2) Noise Filtering. )e interference signals in the image type is determined, and the image processing result is sent to cannot be completely avoided, so these noise signals the memory. must be filtered in a targeted manner. )e image Image processing technology is the core of the tennis signal is video image automatic processing system. )e efficiency of image processing determines the efficiency of automatic f(x, y) � g(x, y) + α(x, y). (28) monitoring. )is article focuses on the key aspects of image processing, such as image segmentation, noise filtering, and In the formula, g(x, y) is the characteristic image; edge detection. α(x, y) is the noise signal in the image. (1) In the process of image automation processing in this Commonly used noise filtering methods include article, the first thing to do is image segmentation. Kalman filtering, median filtering, and so on. )e Image segmentation refers to segmenting the more gray value of the noise signal in the image is 8 Complexity

handle between the continental style and the anti-Eastern style Tennis video Storage (see Figure 5 for the specific handle). It can be seen that the five Imgae input collector device subjects all changed their handling styles when switching from playing to serve, and the “V” in the mouth of the clapping hand changed from the front type to the back type, but each subject Image had a different degree of conversion. output Because the five test subjects in this study are all right- Image handed rackets, in order to facilitate the description of the FPG1 processor left and right sides of the batter’s body in the paper, the RS422 serial “right side” in the text refers to the bat’s “holding side” and port the “left side” is the bat’s “nonholding side.” )rowing the ball vertically helps the hitter avoid too much wrong prediction of the spatial position of the ball, Storage RS422 serial thus greatly improving the accuracy of the shot. When Other system device port throwing the ball, the athlete should ensure that all joints of the upper arm are fully extended, and the relative positions Figure 3: Motion monitoring image system architecture. of the upper arm, shoulder, elbow, and body are directly determined. Moreover, the athlete should throw the ball to the upper right front of the body to provide protection for obviously different from the gray value of the image. the accurate hitting of the ball when preparing to hit the ball )erefore, this paper adopts the median filtering with the right arm swing. In addition, the forward deviation method to filter the signals with abnormal gray value of the position of the ball relative to the body is to conform to in the image to obtain higher quality images. the inertia of the person swinging forward. )erefore, the (3) Edge Detection. In the process of image processing, slight forward shift when throwing the ball is to effectively edge detection plays a very important role in the ensure that the ball can be hit directly above the body when recognition of image feature types. Edge detection is swinging. Table 2 and Figures 6 and 7 show the results of the also a key step to achieve target recognition. Because research the position of throwing shots for the flat strike and the edge image of the target object is obviously the top spin. different from the background, the feature recog- From the overall comparison, each subject’s throws are nition of the target can be effectively performed by different. However, a comparative analysis of each subject’s using edge detection technology. )e edge detection two throwing actions when serving the ball showed that the operators used in this article are Sobel operator and deviation of the throwing in the two horizontal axes (X-axis Robert operator during image processing. )e and Y-axis) in the positive direction of the top spin is larger function of the image operator is to perform accurate than that of the throwing of the flat strike. In other words, it filtering of the edge image, which helps to improve is more appropriate to throw the ball toward the back left of the recognition efficiency of feature types. the batter when a top spin is used. In the throwing and racket phases, the flexion of the 6. Competition Technical Analysis ankle joints, knee and hip joints, and the decrease of the body’s center of gravity are sufficient preparations for the )is paper first obtains the special objects of tennis match explosive force and rapid rebound required for the next technical analysis through data mining technology and vigorous swing. When knees and hips are bent downward, obtains the most representative research objects. After that, the leg muscles are stretched to a certain extent, which this paper specifically analyzes the tennis technology of the induces stretch and launch, which is conducive to the for- research object and names the five research objects A, B, C, mation of explosive force. However, the initial length of the D, and E obtained by data mining in this paper. )e basic musculature varies from person to person, so the optimal conditions are shown in Table 1. angles of the knees and hips that are most conducive to In this paper, two professional digital cameras are used explosive exertion when bending and squatting to lower the to shoot the subject with fixed-point and fixed-focus center of gravity are also different for different people. shooting. Moreover, this paper takes five outstanding pro- Moreover, the results obtained differ among different test fessional graduate students with six years or more of tennis subjects. )e biggest difference is the smallest joint angle experience as the research objects, arranges them to do (Table 3 and Figures 8 and 9). overhand serve techniques and shoot them on the spot. )e In the racket stage, the movement characteristics of the shooting frequency is 240 Hz, the shutter speed is 1/2000 s, left and right shoulder joints are mainly reflected in the the range of the three-dimensional calibration frame is 2.5 m changes in their spatial positions. In the course of the two in diameter, and the angle between the two cameras and the kinds of serve techniques, the vertical height difference test object is 90°, as shown in Figure 4. between the left and right shoulders behaves differently. In terms of handle, the five subjects usually play continental Among them, the change in the height of the shoulder joint style. However, each tester’s handle style changed during the on the nonholding side is that the height of the top spin is serve. A and B use the anti-Eastern style, while C, D, and E use a higher than that of the flat strike, and the average difference Complexity 9

Table 1: Research object information table. Test object Age Height (m) Weight (kg) Practice time (years) Clap hands A 31 1.8 68.7 8.1 Right hand B 32 1.8 73.7 9.1 Right hand C 30 1.8 79.8 7.1 Right hand D 31 1.9 82.8 6.1 Right hand E 29 1.8 80.8 8.1 Right hand

Y Z

X

AB

Figure 4: Experimental recording scene.

spin. In this process, the knee and hip joints also have similar Continental characteristics to the shoulder joints. )e difference between the knee and hip joints in the two technical actions is re- Anti-eastern Eastern way way flected in the increase in the top spin. Table 4 and Figure 10 show the comparison of spatial characteristics at the lowest point of the shoulder joint in the Half western Racket handle racket stage. style From the motion characteristics of the knee, shoulder, and hip joints, we can conclude that the right knee, hip, and shoulder joints have lower flexion compared to the left knee, Western style hip, and shoulder joints when top spin is used. Moreover, the angles of the right knee and hip joints of the top spin are Figure 5: Tennis racket handle. smaller than the angles of the knee and shoulder joints on the side holding the racket of the flat strike. From this, it can be concluded that in racket action, when the top spin is used, is three centimeters. However, the change in the vertical the main joints on the right side are more fully prepared for height of the shoulder joint on the holding side shows that the explosive power of swing than when the flat strike is the height of the top spin is lower than that of the flat strike, used. and the difference is five centimeters on average. Due to the From Table 5 and Figure 11, it is found that although the characteristics of the above two serving techniques, the vertical height of the wrist is not very different, the racket difference in height between the shoulders is more prom- and the ground are almost vertical at the moment of hitting inent, and the average difference is about eight centimeters. when the flat strike is adopted, and the racket is basically Similarly, in terms of the level of space between the left and tilted flat at the moment of hitting when the top spin is right shoulders, when using top spin, it is also greater than adopted. )erefore, the height of the hitting point is different when using flat strike. )e difference in the vertical height of between the former and the latter. the left and right shoulders between the two types of serve is )e height comparison (unit: m) of the hitting point of caused by the increase in the degree of squatting of the knee the flat strike and the top spin is shown in Table 6 and and hip joints on the right side (on the racket side) of the top Figure 12. 10 Complexity

Table 2: Results of the research on the position of throwing shots for the flat strike and the top spin. Research object strike Top spin Flat strike Top spin Flat strike Top spin Flat strike Top spin Flat strike Top spin A 1.67 1.77 7.39 6.88 0.67 1.08 −0.94 −0.21 7.09 6.83 B 1.66 1.73 6.20 6.09 −0.35 0.37 −0.61 0.05 6.00 5.95 C 1.70 1.75 6.54 6.35 −0.22 0.49 0.34 0.43 6.19 6.13 D 1.69 1.75 6.33 6.22 −0.41 0.25 0.55 1.26 6.39 6.30 E 1.71 1.80 6.65 6.63 −0.23 0.65 −0.15 0.36 6.48 6.47

As shown in Table 6 and Figure 12, in tennis, the hitting 8 point refers to the specific spatial position of the ball when it is swung and hitting the ball. In fact, it refers to the relative 7 position of the hitter’s body. When the player hits the ball, 6 the spatial position of each part of the hitter also changes, but only the position of the head is relatively fixed. )erefore, it 5 is more practical to regard the position of the head as a 4 relative reference marking point describing the spatial po- sition of the hitting point. 3 )e three-dimensional distance (unit: m) between the 2 hitting point and the head space position of flat strike is 1 shown in Table 7 and Figure 13. )e three-dimensional distance (unit: m) between the hitting point and the head 0 space position of the top spin is shown in Table 8 and –1 Figure 14. Comparing the two types of serve actions, it can be found –2 Shot height Shot rate X-axis Y-axis Z-axis that in the top spin, a certain degree of deviation of the ball’s (m) (m/s) speed speed speed spatial position is beneficial to the effect of the ball. (m/s) (m/s) (m/s) However, if the degree of deviation is insufficient, the A D topspin effect is not obvious. Due to the limitations of B E various related equipment and other conditions, there is no C way to measure the rotation rate of the ball after being hit in Figure the video. However, we can analyze and judge the specific 6: Statistical diagram of the position of the shot of the flat rotation of the ball by using the subspeed of the racket head strike. in other directions at the moment of batting. )rough the research and analysis of the video data of the five experi- mental testers, it can be concluded that the five experimental 8 testers’ hitting points when using top spin are all deviated to 7 a certain degree to the left and rear than when using flat strike. Compared with the head space position, the more the 6 hitting point deviates to the left, the higher the speed of the 5 top of the racket at the moment of hitting, and the greater the top spin rate of the ball. Conversely, if the hitting point is 4 shifted to the left insufficiently, the top spin rate of the racket 3 will be relatively reduced at the moment of impact, the top spin rate of the ball will decrease, the lateral spin rate will 2 increase, and the lateral spin rate of the ball will increase. In 1 addition, at the moment before hitting the ball, reasonable 0 elbow bending and wrist swing techniques can have a certain effect on the final speed of the racket head. Moreover, the –1 shaking of the wrist makes a significant difference in the –2 speed of the grip point and the speed of the wrist joint. Shot height Shot rate X-axis Y-axis Z-axis Because the grip point is close to the wrist joint point, the (m) (m/s) speed speed speed obvious speed difference must come from the final shaking (m/s) (m/s) (m/s) of the wrist joint. However, the flexion of the elbow and the A D shaking of the wrist should not be too large. Excessive flexion B E of the elbow and wrist joints will cause changes in the flight C path of the ball and ultimately reduce the quality of the serve. Figure 7: Statistical diagram of the position of the shot of the top Moreover, it is especially important that a large swipe of the spin. wrist can easily cause the ball to fall into the net. Complexity 11

Table 3: Comparison table of smallest angles of the knee and hip joints at the racket stage (unit: °). Side knee joint Right knee joint Left hip joint Right hip Flat strike Top spin Flat strike Top spin Flat strike Top spin Flat strike A 85.57 89.29 78.09 74.43 A 85.57 89.29 78.09 B 92.27 94.18 110.58 102.88 B 92.27 94.18 110.58 C 76.21 82.98 103.35 104.95 C 76.21 82.98 103.35 D 100.75 102.84 92.94 85.13 D 100.75 102.84 92.94 E 92.60 86.24 111.27 104.30 E 92.60 86.24 111.27

200

150

100

50

0 A B C D E

Side knee joint Right knee joint Le hip joint Right hip Figure 8: Comparison of the smallest angles of the knee and hip joints at the racket stage of the flat strike.

200

150

100

50

0 A B C D E

Side knee joint Right knee joint Le hip joint Right hip Figure 9: Comparison of the smallest angles of the knee and hip joints at the racket stage of the top spin.

Table 4: Comparison of spatial characteristics at the lowest point of the shoulder joint in racket stage. Left shoulder height Right shoulder height Height difference between )e angle between the shoulders and the (m) (m) shoulders (m) ground (°) Flat strike 1.31 1.16 0.15 36.76 Top spin 1.34 1.11 0.23 43.83 Difference −0.03 0.05 −0.08 −7.07

In tennis, there are differences in the postures of flat of hitting, while the angle between the top spin and the strike and top spin. )e specific difference is the angle ground at the moment of hitting is relatively small. By formed between the flat strike and the top spin. )e flat comparing the angles formed between the two serving strike is almost perpendicular to the ground at the moment methods and the ground in Table 9 and Figure 15, it can be 12 Complexity

50

40

30

20

10

0

–10 Left shoulder Right shoulder Height The angle height (m) height (m) difference between the between shoulder and shoulders (m) the ground (°)

Flat strike Top spin Difference

Figure 10: Comparison of spatial characteristics at the lowest point of the shoulder joint in the racket stage.

Table 5: Comparison of spatial characteristics at the lowest point of the shoulder joint in racket stage. Left knee Right knee Left hip Right hip Left shoulder Right shoulder Right elbow Right wrist Flat strike 0.74 0.77 1.07 1.13 1.46 1.76 2.06 2.27 Top spin 0.75 0.78 1.06 1.14 1.42 1.78 2.04 2.26 Difference −0.01 −0.01 0.01 −0.01 0.04 −0.02 0.02 0.01

2.90 3 2.85 2 2.80

2 2.75 2.70 1 2.65 1 2.60 2.55 0 2.50

2.45

Lehip Leknee

Right hip Right AB DC EAverage

Right knee Right Right wrist Right

Right elbow Right value

Leshoulder Right shoulder Right Flat strike Flat strike Top spin Top spin Figure 12: Comparison diagram of the height of the hitting point Figure 11: Comparison of spatial characteristics of the lowest point of the flat strike and the top spin (unit: m). of the shoulder joint in the racket stage.

Table 6: Comparison table of the height of the hitting point of the Table 7: )e three-dimensional distance between hitting point and flat strike and the top spin (unit: m). head space position of the flat strike (unit: m). ABCDE Average value X-axis Y-axis Z-axis Flat strike 2.69 2.67 2.86 2.67 2.76 2.73 Hitting point 1.11 0.32 1.98 Top spin 2.65 2.61 2.82 2.64 2.72 2.69 Head position 1.27 0.71 1.00 Difference 0.04 0.06 0.04 0.03 0.04 0.04 Spatial distance −0.16 −0.38 0.98 Complexity 13

2.5 Table 9: Statistical table of the angle between the racket and the ground at the moment of hitting the ball (unit: °). 2.0 ABCDE Average value Flat strike 70.46 71.46 71.83 67.40 60.43 68.32 1.5 Top spin 55.37 59.24 54.28 57.96 49.92 55.36 Difference 15.09 12.22 17.55 9.43 10.50 12.96 1.0

80 0.5 70 0.0 60 –0.5 50

–1.0 X-axis Y-axis Z-axis 40

Hitting point 30 Head position Spatial distance 20 Figure 13: Statistical diagram of the three-dimensional distance 10 between the hitting point and the head space position of the flat strike (unit: m). 0 AB DC EAverage value Table 8: )e three-dimensional distance between hitting point and Flat strike head space position of the top spin (unit: m). Top spin

X-axis Y-axis Z-axis Figure 15: Statistical diagram of the angle between the racket and Hitting point 1.30 0.59 1.96 the ground at the moment of hitting the ball (unit: °). Head position 1.24 0.77 0.98 Spatial distance 0.06 −0.18 0.98 the moment of hitting and the ground plays a decisive role in 2.5 determining the vertical height of the hitting point. Using machine learning to analyze the characteristics of the tennis movement process can effectively improve the 2.0 effect of feature extraction and discover the problems of athletes in competition and training. Data feature extraction 1.5 is an effective way to reflect the actual situation through experiments, and it is also an inevitable development of sports direction. 1.0

0.5 7. Conclusion

0.0 One of the main focuses of machine learning research is to improve the generalization accuracy and efficiency of model prediction, but in many applications, it lacks the ability to –0.5 transform prediction results into action knowledge. )is X-axis Y-axis Z-axis paper first introduces the research background and the Hitting point current research status at home and abroad and gives a Head position concise overview of data mining and knowledge extraction Spatial distance and explains the importance of the study of action Figure 14: Statistical diagram of the three-dimensional distance knowledge extraction methods from data mining. )is paper between the hitting point and the head space position of the top proposes an action knowledge extraction method that ex- spin (unit: m). tracts high prediction accuracy and high efficiency decision- making action from data and applies this method to the concluded that there is no significant difference in the technical analysis of tennis. vertical height of the wrist joints when the two serving With the tremendous advances in data science, the ul- methods hit the ball. )erefore, the angle formed between timate goal of extracting patterns from data is to facilitate 14 Complexity decision-making. It is foreseeable that machine learning experiences,” Athletic Training Education Journal, vol. 10, models will become part of making rational decisions in no. 2, pp. 103–112, 2015. large-scale AI systems. )ese modes will be crucial to the [9] O. V. Ivashchenko, O. O. Kapkan, and O. O. Kapkan, support of the action. )is research uses SIMI Motion “Simulation of process of 14-15 years old girls’ training of light human motion video processing software to display the athletic and gymnastic exercises,” Pedagogics, Psychology, changing form of each action during the serve in a digital Medical-Biological Problems of Physical Training and Sports, vol. 19, no. 8, pp. 32–39, 2015. way. Moreover, this research processes the acquired data, [10] P. D. Owen and N. King, “Competitive balance measures in analyzes and summarizes the reasonable data obtained by sports leagues: the effects of variation in season length,” logic and reasoning, and analyzes the data in combination Economic Inquiry, vol. 53, no. 1, pp. 731–744, 2015. with the model constructed in this paper. )e research [11] J. Yang, “)e simulation of table tennis during the course of results show that the performance of the model constructed sports,” Caribbean Journal of Science, vol. 52, no. 4, in this paper is good. pp. 1561–1564, 2019. )is article uses experimental research methods to ex- [12] M. Bulat, N. Korkmaz Can, Y. Z. Arslan, and W. Herzog, tract and analyze the action characteristics of tennis. )e “Musculoskeletal simulation tools for understanding mech- model built in this article is only limited to the experimental anisms of lower-limb sports injuries,” Current Sports Medicine stage, and the follow-up needs to be studied from the Reports, vol. 18, no. 6, pp. 210–216, 2019. perspective of actual competition and training, and the [13] A. Lopatiev, O. Ivashchenko, O. Khudolii et al., “Systemic approach and mathematical modeling in physical education amount of data in this article is relatively small. In future and sports,” Journal of Physical Education and Sport (JPES), research, it is needed to use a lot of data for analysis. vol. 17, no. 1, pp. 146–155, 2017. [14] N. Bennour, “Teaching practices and student action in Data Availability physical education classes: perspectives for teacher educa- tion,” Creative Education, vol. 6, no. 10, pp. 934–944, 2015. )e raw/processed data required to reproduce these findings [15] D. Sghaier, M. S. Elandoulsi, and A. Bouassida, “Physical cannot be shared at this time as the data also form part of an education teacher’s training in swimming under the joint ongoing study. didactic action,” Creative Education, vol. 6, no. 22, pp. 2433–2437, 2015. [16] N. Cutforth and E. S. Belansky, “A community-engaged Conflicts of Interest approach to translating research into practice: a physical education story,” Progress in Community Health Partnerships: )e authors declare that they have no conflicts of interest. Research, Education, and Action, vol. 9, no. 4, pp. 571–582, 2015. References [17] E.-R. Hwang and T.-Y. Kim, “Intensification of the education of public health, hygiene, and martial arts during the Japanese [1] P. S. Glazier and S. Mehdizadeh, “Challenging conventional colonial period (1937–1945),” Journal of Exercise Rehabili- paradigms in applied sports biomechanics research,” Sports tation, vol. 14, no. 2, pp. 160–167, 2018. Medicine, vol. 49, no. 2, pp. 171–176, 2019. [18] D. Murase, K. K. Yokoyama, and Y. Y. YamamotoHasegawa, [2] T. Mullen, J. Highton, and C. Twist, “)e internal and external “Baseball catching patterns differ according to task con- responses to a forward-specific rugby league simulation straints,” Advances in Physical Education, vol. 6, no. 3, protocol performed with and without physical contact,” In- pp. 151–157, 2016. ternational Journal of Sports Physiology and Performance, [19] J. T. Spector and M. Lieblich, “Letter regarding “Comparison vol. 10, no. 6, pp. 746–753, 2015. between low-cost marker-less and high-end marker-based [3] L. G. Appelbaum and G. Erickson, “Sports vision training: a motion capture systems for the computer-aided assessment of review of the state-of-the-art in digital training techniques,” working ergonomics” by Patrizi et al. and research repro- International Review of Sport and Exercise Psychology, vol. 11, ducibility,” Ergonomics, vol. 60, no. 4, pp. 597-598, 2017. no. 1, pp. 160–189, 2018. [20] F. S´a,A. Marques, N. B. F. Rocha, M. J. Trigueiro, C. Campos, [4] R. Rein and D. Memmert, “Big data and tactical analysis in and J. Schroder,¨ “Kinematic parameters of throwing perfor- elite soccer: future challenges and opportunities for sports mance in patients with schizophrenia using a markerless science,” SpringerPlus, vol. 5, no. 1, pp. 1–13, 2016. motion capture system,” Somatosensory & Motor Research, [5] K. R. Ridderinkhof and M. Brass, “How kinesthetic motor vol. 32, no. 2, pp. 77–86, 2015. imagery works: a predictive-processing theory of visualization [21] M.-K. Kim, T. Y. Kim, and J. Lyou, “Performance im- in sports and motor expertise,” Journal of Physiology-Paris, provement of an AHRS for motion capture,” Journal of In- vol. 109, no. 1–3, pp. 53–63, 2015. stitute of Control, Robotics and Systems, vol. 21, no. 12, [6] D. Mulligan, K. R. Lohse, and N. J. Hodges, “An action-in- pp. 1167–1172, 2015. congruent secondary task modulates prediction accuracy in [22] Z.-M. Zhou and Z.-W. Chen, “A survey of motion capture experienced performers: evidence for motor simulation,” data earning as high dimensional time series,” International Psychological Research, vol. 80, no. 4, pp. 496–509, 2016. Journal of Multimedia and Ubiquitous Engineering, vol. 10, [7] O. M. Khudolii, O. V. Ivashchenko, S. S. Iermakov et al., no. 9, pp. 17–30, 2015. “Computer simulation of junior gymnasts’ training process,” [23] M. A. Khan, “Multiresolution coding of motion capture data Science of Gymnastics Journal, vol. 8, no. 3, pp. 215–228, 2016. for real-time multimedia applications,” Multimedia Tools & [8] K. A. Tivener and D. S. Gloe, “)e effect of high-fidelity Applications, vol. 76, no. 15, pp. 1–16, 2016. cardiopulmonary resuscitation (CPR) simulation on athletic [24] A. Puupponen, B. T. Wainio, and T. Jantunen, “Head training student knowledge, confidence, emotions, and movements in Finnish Sign Language on the basis of Motion Complexity 15

Capture data: a study of the form and function of nods, nodding, head thrusts, and head pulls,” Sign Language and Linguistics, vol. 18, no. 1, pp. 41–89, 2015. [25] R. Giannetti, A. Petrella, J. Bach et al., “In vivo bone position measurement using high-frequency ultrasound validated with 3-D optical motion capture systems: a feasibility study,” Journal of Medical & Biological Engineering, vol. 37, no. 7, pp. 1–8, 2017. [26] H. Zhang, S. L. Wang, S. Chu, S. Chen, H. Meng, and G. Liu, “Application of optical motion capture technology in power safety entitative simulation training system,” Optics and Photonics Journal, vol. 6, no. 8, pp. 155–163, 2016.